The open source model release velocity is unsustainable. Here's why.
I counted 142 models released on Hugging Face in February 2025 alone. That's 5 per day. Downloads are up but download-per-model is falling. The attention pie is finite. I think a shakeout is coming.
I counted every notable model uploaded to Hugging Face in February 2025.
142 models. In 28 days. That's 5.07 models per day.
The number sounds exciting until you look at the download data.
The release velocity trend
| Quarter | Notable model releases | Models per day | |---------|----------------------|---------------| | Q1 2023 | 31 | 0.34 | | Q2 2023 | 38 | 0.42 | | Q3 2023 | 42 | 0.46 | | Q4 2023 | 51 | 0.55 | | Q1 2024 | 58 | 0.64 | | Q2 2024 | 63 | 0.69 | | Q3 2024 | 72 | 0.78 | | Q4 2024 | 89 | 0.97 | | Jan 2025 | 127 (monthly) | 4.10 | | Feb 2025 | 142 (monthly) | 5.07 |
Sources: Hugging Face model hub, my tracking spreadsheet. "Notable" = models with technical reports or from established labs.
The jump from Q4 2024 to early 2025 is dramatic. We went from roughly 1 model per day to 5 per day in three months. DeepSeek R1 and its distilled variants kicked off a wave of reasoning model releases. Everyone with a GPU cluster and a RL recipe started publishing.
Downloads are up, but downloads per model are down
| Quarter | Total HF downloads (top 200 models) | Avg downloads per model | |---------|--------------------------------------|------------------------| | Q1 2024 | 2.1B | 10.5M | | Q2 2024 | 2.8B | 9.3M | | Q3 2024 | 3.4B | 8.1M | | Q4 2024 | 4.1B | 6.7M | | Jan-Feb 2025 | 3.2B (annualized ~19B) | 5.1M |
Sources: Hugging Face download counters, my analysis.
Total downloads keep going up. The market is growing. But the average downloads per model peaked in Q1 2024 and has been declining since. There's more supply than demand can absorb.
Who's releasing what
| Organization | Models in Feb 2025 | Notable | |-------------|-------------------|---------| | DeepSeek | 8 | R1 distilled variants | | Meta AI | 3 | Llama 3.3 variants | | Alibaba/Qwen | 11 | Qwen 2.5 fine-tunes | | Mistral AI | 4 | Mistral Small, Codestral | | Community fine-tunes | 86 | Various merges, LoRA | | Other labs | 30 | Yi, Baichuan, InternLM, etc. |
Sources: Hugging Face model hub, February 2025.
Community fine-tunes account for 60% of releases. Most of these are LoRA adapters, model merges, or quantized versions of existing models. They're not genuinely new models. They're variations on a theme.
Of the 142 releases, I'd call maybe 20-25 of them genuinely new (distinct architecture, training data, or approach). The rest are derivatives.
The attention problem
Here's what worries me. There are now more good models than any single person or team can evaluate. I track this stuff obsessively, and even I can't keep up.
I tested 8 new reasoning models last month. That took 40+ hours of evaluation time. There were at least 15 more I didn't get to.
When benchmarks can't keep up with releases, the industry defaults to brand recognition. People use Llama because it's Meta. They use Mistral because it's Mistral. Genuinely better models from smaller labs get ignored.
My prediction
The current release velocity is unsustainable. I expect a shakeout in the next 6-12 months:
| Prediction | Rationale | |-----------|-----------| | 30% fewer labs releasing models by end of 2025 | Compute costs + diminishing returns on attention | | Community fine-tunes plateau | Base models converging means less room for improvement | | 3-5 dominant open source families | Llama, Qwen, Mistral, DeepSeek, plus maybe one wild card | | More focus on deployment, less on training | The "model release" phase shifts to "model deployment" phase |
The open source AI movement isn't slowing down. But it's about to consolidate around fewer, larger players. The era of "everyone releases a model" is peaking right now.
Five models per day. I can barely keep my spreadsheet updated. Something's gotta give.
If you found this interesting, you might also like:
- Hugging Face just hit 10,000 models. Here's what the model zoo looks like.
- Open source AI is having a moment. Here are the download numbers.
- Wait, Stable Diffusion has HOW many forks? The open source explosion in numbers.
- The GPT-3 API waitlist is 6 months long. Here's what the early data looks like.
- I counted every AI startup that raised money in Q1 2021. The numbers are strange.
-- dataku